udpate ensemble example notebook (#1996)

examples/19-EnsembleModel-examples.ipynb

@@ -545,7 +545,9 @@
     "\n",
     "When predicting a number of values greater than their `output_chunk_length`, `GlobalForecastingModels` rely on auto-regression (use their own output as input) to forecast values far in the future. However, the quality of the forecasts can considerably decrease as the predicted timestamp get further from the end of the observations. During `RegressionEnsemble`'s regression model training, the forecasting models generate forecasts for timestamps where the ground truth is actually known and available, making it possible to use `historical_forecasts` instead of `predict()`.\n",
     "\n",
-    "**Note**: In order to ensure that `regression_train_n_points` are generated, the `output_chunk_length` of all the forecasting models must be multiple of `regression_train_n_points`. If the forecasting models expect covariates, they should extend far enough into the future. If these two conditions are not met, a warning message will be displayed and the regression model of the ensemble might be trained with less points than specified."
+    "This can be activated with `train_using_historical_forecasts=True`.\n",
+    "\n",
+    "Under the hood, the ensemble model will trigger historical forecasting for each model with `forecast_horizon=model.output_chunk_length`, `stride=model.output_chunk_length`, `last_points_only=False`, and `overlap_end=False` to predict the last `regression_train_n_points` points of the target series."
    ]
   },
   {
@@ -594,7 +596,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "As expected, using historical forecast with the forecasting models to the train the regression model produce better forecasts.\n",
+    "As expected, using historical forecasts with the forecasting models to the train the regression model produces better forecasts.\n",
     "\n",
     "### Probabilistic regression ensemble\n",
     "\n",